the idea
if agents handle money, they need credit scores. not for borrowing — for trust. before you give an agent a $10k/month API budget, you want to know: does it spend predictably? does it respect limits? has it ever tried to move money after its context window was poisoned?
how it works
mnemopay logs every agent payment to a merkle chain: timestamp, amount, recipient, reasoning snippet, approval hash. that's the raw data.
i'm piping those logs into a typescript anomaly detection library (github has 40+ options now). the detector learns normal patterns — "this agent spends $200–$400/day on AWS, always between 9am–5pm UTC, never to new vendors without a reasoning chain longer than 80 tokens."
when an agent tries something weird — $8k to a new vendor at 3am with a 12-token reasoning chain — the anomaly score spikes. if it crosses a threshold, FiscalGate (our two-phase commit layer) holds the payment for human review.
why Agent FICO matters
right now agents are single-tenant. one company, one agent, one budget. but agents will job-hop — an agent that managed marketing spend at company A applies to do procurement at company B. company B wants references. Agent FICO is the reference: a portable reputation score derived from audited payment history.
the TS anomaly libs are good enough today. the missing piece was the tamper-evident log. that's what mnemopay ships.
Top comments (0)